mlelcd: Compute the maximum likelihood estimator of a log-concave density

Description

Uses Shor's $r$-algorithm to compute the maximum likelihood estimator of a log-concave density based on an i.i.d. sample. The estimator is uniquely determined by its value at the data points. The output is an object of class "LogConcDEAD" which contains all the information needed to plot the estimator using the plot method, or to evaluate it using the function dlcd.

Usage

mlelcd(x, w=rep(1/nrow(x),nrow(x)), y=initialy(x),
  verbose=-1, alpha=5, c=1, sigmatol=10^-8, integraltol=10^-4,
  ytol=10^-4, stepscale=5.1, stepscale2=2, stepscale3=1.5,
  stepscale4=1.05, desiredsize=3.3, Jtol=0.001, chtol=10^-6)

Arguments

Data in $R^d$, in the form of an $n \times d$ numeric matrix

Vector of weights $w_i$ such that the computed estimator maximizes $$\sum_{i=1}^n w_i \log f(x_i)$$

Vector giving starting point for the $r$-algorithm. If none given, a kernel estimate is used.

verbose

-1: (default) prints nothing
0: prints warning messages
$n>0$: prints summary information every$n$iterations

alpha

Scalar parameter for SolvOpt

Scalar giving starting step size

sigmatol

Real-valued scalar giving one of the stopping criteria: Relative change in $\sigma$ must be below sigmatol for algorithm to terminate. (See Details)

ytol

Real-valued scalar giving on of the stopping criteria: Relative change in $y$ must be below ytol for algorithm to terminate. (See Details)

integraltol

Real-valued scalar giving one of the stopping criteria: $| 1 - \exp(\bar{h}_y) |$ must be below integraltol for algorithm to terminate. (See Details)

stepscale, stepscale2, stepscale3, stepscale4, desiredsize

Scalar parameters for SolvOpt. Changing these is not recommended.

Jtol

Parameter controlling when Taylor expansion is used in computing the function $\sigma$

chtol

Parameter controlling convex hull computations

Value

An object of class "LogConcDEAD", with the following components:
xData copied from input (may be reordered)
wweights copied from input (may be reordered)
logMLEvector of the log of the maximum likelihood estimate, evaluated at the observation points
NumberOfEvaluationsVector containing the number of steps, number of function evaluations, and number of subgradient evaluations. If the SolvOpt algorithm fails, the first component will be an error code $(<0)$.< description="">
MinSigmaReal-valued scalar giving minimum value of the objective function
bmatrix (see Details)
betavector (see Details)
triangmatrix containing final triangulation of the convex hull of the data
vertsmatrix containing details of triangulation for use in dlcd
vertsoffsetmatrix containing details of triangulation for use in dlcd
chullVector containing vertices of faces of the convex hull of the data
outnormmatrix where each row is an outward pointing normal vectors for the faces of the convex hull of the data. The number of vectors depends on the number of faces of the convex hull.
outoffsetmatrix where each row is a point on a face of the convex hull of the data. The number of vectors depends on the number of faces of the convex hull.

Details

The log-concave maximum likelihood density estimator based on data $X_1, \ldots, X_n$ is the function that maximizes $$\sum_{i=1}^n w_i \log f(X_i)$$ subject to the constraint that $f$ is log-concave. For i.i.d.~data, the weights $w_i$ should be $\frac{1}{n}$ for each $i$. This is a function of the form $\bar{h}_y$ for some $y \in R^n$, where $$\bar{h}_y(x) = \inf \lbrace h(x) \colon h \textrm{ concave }, h(x_i) \geq y_i \textrm{ for } i = 1, \ldots, n \rbrace.$$

Functions of this form may equivalently be specified by dividing $C_n$, the convex hull of the data, into simplices $C_j$ for $j \in J$ (triangles in 2d, tetrahedra in 3d etc), and setting $$f(x) = \exp{b_j^T x - \beta_j}$$ for $x \in C_j$, and $f(x) = 0$ for $x \notin C_n$. This function uses Shor's $r$-algorithm (an iterative subgradient-based procedure) to minimize over vectors $y$ in $R^n$ the function $$\sigma(y) = -\frac{1}{n} \sum_{i=1}^n y_i + \int \exp(\bar{h}_y(x)) \, dx.$$ This is equivalent to finding the log-concave maximum likelihood estimator, as demonstrated in Cule, Samworth and Stewart (2008). An implementation of Shor's $r$-algorithm based on SolvOpt is used.

Computing $\sigma$ makes use of the qhull library. Code from this C-based library is copied here as it is not currently possible to use compiled code from another library. For points not in general position, this requires a Taylor expansion of $\sigma$, discussed in Cule and D"umbgen (2008).

References

Barber, C.B., Dobkin, D.P., and Huhdanpaa, H.T. (1996) The Quickhull algorithm for convex hulls ACM Trans. on Mathematical Software, 22(4) p.469-483 http://www.qhull.org

Cule, M. L. and D"umbgen, L. (2008) On an auxiliary function for log-density estimation, University of Bern technical report. http://arxiv.org/abs/0807.4719

Cule, M. L., Samworth, R. J., and Stewart, M. I. (2010) Maximum likelihood estimation of a log-concave density, Journal of the Royal Statistical Society, Series B, 72(5) p.545-607.

Kappel, F. and Kuntsevich, A. V. (2000) An implementation of Shor's r-algorithm Computational Optimization and Applications 15 http://www.uni-graz.at/imawww/kuntsevich/solvopt/

Shor, N. Z. (1985) Minimization methods for nondifferentiable functions Springer-Verlag

Examples

Run this code

## Some simple normal data, and a few plots

x <- matrix(rnorm(200),ncol=2)
lcd <- mlelcd(x)
g <- interplcd(lcd)
par(mfrow=c(2,2), ask=TRUE)
plot(lcd, g=g, type="c")
plot(lcd, g=g, type="c", uselog=TRUE)
plot(lcd, g=g, type="i")
plot(lcd, g=g, type="i", uselog=TRUE)

## Some plots of marginal estimates
par(mfrow=c(1,1))
g.marg1 <- interpmarglcd(lcd, marg=1)
g.marg2 <- interpmarglcd(lcd, marg=2)
plot(lcd, marg=1, g.marg=g.marg1)
plot(lcd, marg=2, g.marg=g.marg2) 

## generate some points from the fitted density
## via independent rejection sampling
generated1 <- rlcd(100, lcd)
genmean1 <- colMeans(generated1)
## via Metropolis-Hastings algorithm
generated2 <- rlcd(100, lcd, "MH")
genmean2 <- colMeans(generated2)

## evaluate the fitted density
mypoint <- c(0, 0)
dlcd(mypoint, lcd, uselog=FALSE)
mypoint <- c(10, 0)
dlcd(mypoint, lcd, uselog=FALSE)

## evaluate the marginal density
dmarglcd(0, lcd, marg=1)
dmarglcd(1, lcd, marg=2)

## evaluate the covariance matrix of the fitted density
covariance <- cov.LogConcDEAD(lcd)

## find the hat matrix for the smoothed log-concave that
## matches empirical mean and covariance
A <- hatA(lcd)

## evaluate the fitted smoothed log-concave density
mypoint <- c(0, 0)
dslcd(mypoint, lcd, A)
mypoint <- c(10, 0)
dslcd(mypoint, lcd, A)

## generate some points from the fitted smoothed log-concave density
generated <- rslcd(100, lcd, A)

Run the code above in your browser using DataLab